Non-commutative Crepant Resolutions (NCCRs)

• NCCRs are non-commutative analogues of crepant resolutions, defined via endomorphism rings of reflexive modules with finite global dimension and Cohen–Macaulay properties.
• They are constructed using tilting bundles and derived equivalences, linking geometric crepant resolutions to algebraic and combinatorial models like dimer configurations.
• NCCRs provide a powerful framework bridging birational geometry, representation theory, and toric techniques, leading to new insights in singularity resolution and mutation theory.

Non-commutative crepant resolutions (NCCRs) are non-commutative analogues of crepant resolutions in algebraic geometry. In the standard formulation for a normal Gorenstein ring $R$, one seeks an algebra of the form

$\Lambda=\End_R(M),$

where $M$ is a nonzero reflexive $R$-module, $\Lambda$ has finite global dimension, and $\Lambda$ is Cohen–Macaulay as an $R$-module; in geometric situations such algebras are expected to recover, or be recovered from, tilting objects on crepant resolutions and their derived categories (Leuschke, 2011). The theory connects birational geometry, tilting theory, the McKay correspondence, toric geometry, dimer models, and representation theory, and it now includes quotient, toric, determinantal, nilpotent-orbit, and del Pezzo examples, together with mutation and wall-crossing formalisms (Bergh, 2022).

1. Formal definitions and homological formulations

A widely used definition, going back to Van den Bergh and adopted in several later treatments, starts with a normal Gorenstein $k$-algebra $R$ and a reflexive $R$-module $\Lambda=\End_R(M),$0, and requires that

$\Lambda=\End_R(M),$1

have finite global dimension and be Cohen–Macaulay over $\Lambda=\End_R(M),$2 (Špenko et al., 2017). Expository accounts often package this in the language of non-singular $\Lambda=\End_R(M),$3-orders: $\Lambda=\End_R(M),$4 is a non-singular $\Lambda=\End_R(M),$5-order if it is finitely generated and Cohen–Macaulay over $\Lambda=\End_R(M),$6 and satisfies

$\Lambda=\End_R(M),$7

and an NCCR is then an endomorphism ring $\Lambda=\End_R(M),$8 with this property (Iyama et al., 2015).

A more elaborate formulation emphasizes birationality and symmetry. In Van den Bergh’s “scenes from categorical geometry,” an NCCR is an $\Lambda=\End_R(M),$9-algebra $M$0 satisfying: birationality $M$1, the order property, finite global dimension, and the symmetry condition

$M$2

as an $M$3-bimodule; equivalently, $M$4 is $M$5-Calabi–Yau in the derived sense (Leuschke, 2011). In the Gorenstein case, this symmetry is the non-commutative version of crepancy.

The terminology broadens outside the normal Gorenstein setting. Survey and foundational papers distinguish non-commutative resolutions, twisted NCCRs, and generalized NCCRs. One overview defines a twisted non-commutative resolution as a reflexive $M$6-algebra that is Azumaya in codimension one and has finite global dimension, and calls it a twisted NCCR when it is also Cohen–Macaulay over $M$7; the untwisted case is the special case $M$8 (Bergh, 2022). Dao–Faber–Ingalls propose a notion over arbitrary commutative noetherian rings: a torsion-free module $M$9 with full support gives an NCCR if $R$0 is a non-singular $R$1-order, a definition that recovers the usual one for normal Gorenstein domains with reflexive $R$2 (Dao et al., 2014).

These formulations are not identical in general. This suggests that “NCCR” is best understood as a family of closely related homological conditions whose common core is finite global dimension together with a crepant or order-theoretic compatibility with the center.

2. Geometric origin in crepant resolutions, tilting theory, and McKay-type correspondences

The geometric source of NCCRs is the observation that a tilting bundle on a crepant resolution produces a non-commutative algebra with the expected homological behavior. If $R$3 is a crepant resolution and $R$4 is a tilting bundle on $R$5, then

$R$6

is an NCCR of $R$7 (Špenko et al., 2017). In the same spirit, when a smooth toric Deligne–Mumford stack or a stacky crepant resolution carries a split tilting bundle, its endomorphism algebra gives a toric NCCR (Špenko et al., 2017).

This construction is closely tied to derived equivalence. Standard tilting theory gives

$R$8

when $R$9 is tilting (Leuschke, 2011). A survey formulation states that if $\Lambda$0 admits a projective crepant resolution $\Lambda$1 whose fibers have dimension $\Lambda$2, then $\Lambda$3 has an NCCR, and conversely, in dimension $\Lambda$4, if $\Lambda$5 has an NCCR $\Lambda$6, then there is a projective crepant resolution $\Lambda$7 with

$\Lambda$8

(Bergh, 2022). Van den Bergh’s threefold results go further: if $\Lambda$9 has a crepant resolution $\Lambda$0 with one-dimensional fibers, then $\Lambda$1 is an NCCR, and in dimension $\Lambda$2 all geometric and non-commutative crepant resolutions of a Gorenstein terminal $\Lambda$3-fold are derived-equivalent (Leuschke, 2011).

The prototypical example is the McKay correspondence. For $\Lambda$4, one has $\Lambda$5, the minimal resolution $\Lambda$6, a tilting tautological bundle on $\Lambda$7, and

$\Lambda$8

which is the standard NCCR (Leuschke, 2011). More generally, for a finite subgroup $\Lambda$9 in characteristic prime to $R$0, the skew-group algebra $R$1 is a natural NCCR of $R$2 when $R$3 has no reflections (Bergh, 2022).

These constructions explain why NCCRs occupy a central place in the non-commutative Bondal–Orlov program: they behave as derived models of crepant birational geometry, especially in dimensions $R$4 and $R$5.

3. Existence theorems and major families

The existence theory begins with quotient singularities. Beyond finite groups, Špenko–Van den Bergh show that quotient singularities for arbitrary reductive groups always have non-commutative resolutions in an appropriate sense, and they exhibit a large class with twisted NCCRs. Their methods are algebraic, do not depend on knowing a commutative resolution, and yield previously unknown twisted NCCRs for determinantal varieties of symmetric and skew-symmetric matrices (Špenko et al., 2015).

Toric singularities form the most extensively developed class. For an abelian reductive group $R$6 and a generic unimodular representation $R$7, Špenko–Van den Bergh prove that if

$R$8

then the invariant ring $R$9 admits an NCCR. This criterion recovers Broomhead’s theorem that every three-dimensional Gorenstein affine toric singularity admits an NCCR, and it also produces a four-dimensional Gorenstein toric singularity with no toric NCCR but with a non-toric NCCR (Špenko et al., 2017). In a companion paper, the same authors give an alternative proof of the three-dimensional toric existence theorem by constructing a split tilting bundle on a stacky toric crepant resolution using standard toric methods rather than dimer models (Špenko et al., 2017).

Several specialized toric families admit more explicit constructions. Gorenstein Hibi rings with class group $k$0 have splitting NCCRs (Nakajima, 2018). For Segre products of polynomial rings, viewed as Hibi rings, one can classify conic divisorial ideals and then construct a splitting NCCR from a finite set of conic rank-one reflexive modules, after which mutation produces further NCCRs (Higashitani et al., 2017). For Gorenstein toric singularities with divisor class group of rank one, toric NCCRs are classified by non-trivial upper sets in a quotient of the divisor class group equipped with a partial order, and all such toric NCCRs are connected by iterated Iyama–Wemyss mutations (Tomonaga, 30 Oct 2025). More recent toric work proves existence for affine toric Gorenstein varieties associated to cones over reflexive polytopes with at most $k$1 vertices by combining toric Deligne–Mumford stacks, exceptional collections, and a generalization of the Špenko–Van den Bergh and Iyama–Wemyss frameworks (Malter et al., 15 Sep 2025).

Two further developments sharpen the higher-dimensional toric picture. Malter proves necessary and sufficient conditions for an incomplete sum of conic modules to give an NC(C)R, reduces existence questions for such endomorphism algebras to the torsion-free class-group case, and classifies the almost simplicial Gorenstein cones that admit NCCRs via endomorphism algebras of conic modules (Malter, 25 Mar 2026). Malter–Sheshmani also show that toric NCCRs descend along lattice-equivalent faces, and derive new short proofs of the existence of toric NCCRs for simplicial and almost simplicial affine toric Gorenstein algebras (Malter et al., 25 Feb 2026).

Outside toric geometry, Hara constructs an NCCR of the minimal nilpotent orbit closure of type $k$2, identifies it with the path algebra of the double Beilinson quiver with relations, and reconstructs the two crepant resolutions as moduli spaces of quiver representations (Hara, 2017). For anticanonical cones over del Pezzo surfaces, every NCCR arises from a geometric helix on the underlying del Pezzo surface (Nordskova et al., 13 Apr 2026).

4. Special classes: toric, splitting, steady, semi-steady, and nonnoetherian variants

A large part of the structure theory concerns NCCRs built from rank-one reflexive modules. In the toric setting, one often writes

$k$3

and calls $k$4 a toric NCCR (Špenko et al., 2018). When $k$5 is a direct sum of rank-one reflexive modules, Iyama–Nakajima call the NCCR splitting; when, in addition, $k$6 is a generator and $k$7, they call it steady (Iyama et al., 2015).

These extra conditions have strong classification consequences. Iyama–Nakajima prove that a singularity has a steady splitting NCCR if and only if it is a quotient singularity by a finite abelian group (Iyama et al., 2015). In the toric threefold case, this means that steady splitting NCCRs single out finite abelian quotient singularities among Gorenstein toric singularities. They also show, for complete local $k$8-dimensional Cohen–Macaulay normal domains over an algebraically closed field of characteristic zero, that the existence of a steady splitting NCCR, a unique basic splitting NCCR, a strongly graded regular local cover, and finiteness properties of the class group are equivalent formulations in the quotient-singularity case (Iyama et al., 2015).

Dimer models provide a complementary realization of splitting NCCRs. A consistent dimer model on the two-torus gives a splitting NCCR of a three-dimensional complete local Gorenstein toric singularity (Nakajima, 2016). Within this class, Nakajima introduces semi-steady NCCRs, defined by the condition that for each indecomposable summand $k$9, the module $R$0 lies in $R$1. The main classification states that a consistent dimer model gives a semi-steady NCCR if and only if it is homotopy equivalent to a regular dimer; steady corresponds exactly to the regular hexagonal tiling, while semi-steady but non-steady corresponds to the square tiling (Nakajima, 2016).

A different generalization appears in the nonnoetherian dimer setting. Beil studies nonnoetherian homotopy dimer algebras as tiled matrix rings and defines nonnoetherian analogues of non-commutative desingularizations and NCCRs. Under cyclic contraction hypotheses, the localization $R$2 is a noncommutative desingularization of its nonnoetherian center, and under an additional coprimeness condition on arrows it becomes a nonnoetherian NCCR of the form

$R$3

with $R$4 reflexive (Beil, 2016).

The proliferation of these variants shows that the basic endomorphism-ring definition supports a fine internal taxonomy. This suggests that “splitting,” “steady,” “semi-steady,” and related adjectives are not merely refinements of taste, but detect rigid geometric and combinatorial features of the underlying singularity.

5. Mutations, wall-crossing, and derived-equivalence phenomena

Mutation theory is one of the central dynamical structures on the set of NCCRs. In the toric setting, Špenko–Van den Bergh prove that for a unimodular, generic, weakly symmetric torus representation $R$5, every toric NCCR of $R$6 is derived equivalent to a fixed Deligne–Mumford GIT quotient stack $R$7. This provides evidence for a non-commutative Bondal–Orlov conjecture: all toric NCCRs of the same affine GIT quotient are derived equivalent (Špenko et al., 2018).

Hara–Hirano relate these derived equivalences to wall-crossing in geometric invariant theory. For generic quasi-symmetric representations of a connected reductive group, they show that equivalences between magic windows corresponding to wall-crossings coincide with derived equivalences between NCCRs induced by tilting modules, and that the relevant tilting modules are produced by exchanges of modules. When the group is a torus, these exchanges are iterated Iyama–Wemyss mutations (Hara et al., 2023). The same paper also uses noncommutative matrix factorizations to obtain an action of

$R$8

on the derived category of a Calabi–Yau complete intersection in a weighted projective space (Hara et al., 2023).

Specific families admit more rigid connectivity results. For toric singularities with divisor class group of rank one, the upper-set classification is accompanied by a proof that all toric NCCRs are connected by iterated Iyama–Wemyss mutations (Tomonaga, 30 Oct 2025). For anticanonical cones over del Pezzo surfaces, every NCCR arises from a geometric helix, and all such geometric helices are connected by mutations, up to tensoring by line bundles and shifts (Nordskova et al., 13 Apr 2026). In a symplectic example, Hara shows that for the minimal nilpotent orbit closure of type $R$9, a certain “multi-mutation” is a composition of Iyama–Wemyss mutations, and the $R$0-twist on the derived category of one crepant resolution corresponds to this non-commutative operation (Hara, 2017).

These results reinforce a central principle of the subject: NCCRs of a fixed singularity typically form a mutation-connected and derived-equivalent family, mirroring the way commutative crepant resolutions are connected by flops.

6. Obstructions, generalizations, and current directions

NCCR existence is constrained by strong necessary conditions. Stafford–Van den Bergh show that homologically homogeneous algebras force the center to have rational singularities in characteristic zero, and survey accounts state that if $R$1 admits a twisted NCCR then $R$2 has rational singularities (Leuschke, 2011). At the same time, existence of a commutative crepant resolution does not guarantee existence of an NCCR: the overview literature notes examples, due to Dao, of factorial hypersurface singularities of dimension $R$3 with crepant resolutions but no NCCRs (Bergh, 2022).

Definitions also become subtle outside the Gorenstein framework. Van den Bergh’s original definition was stated for Gorenstein rings, but representation-theoretic treatments often work for Cohen–Macaulay rings without the Gorenstein hypothesis, phrasing an NCCR as an endomorphism ring that is a non-singular $R$4-order (Iyama et al., 2015). Survey papers emphasize that current definitions fail to cover canonical but non-Gorenstein rings in a fully satisfactory way, so the symmetry condition must be relaxed or reinterpreted (Leuschke, 2011). Dao–Faber–Ingalls address a different limitation by proposing a notion of NCCR over arbitrary commutative noetherian rings, while also proving, for example, that if $R$5 is reduced and $R$6 is homologically homogeneous, then the center of $R$7 is the normalization $R$8 (Dao et al., 2014).

Open problems are especially visible in higher-dimensional toric geometry. One survey formulates the conjecture that every affine Gorenstein toric variety admits an NCCR; this is proved in dimension $R$9 but open in higher dimension (Bergh, 2022). Earlier toric work asks whether the bound

$\Lambda=\End_R(M),$00

is necessary in some sense, and whether more subtle geometric or combinatorial invariants are needed in dimension $\Lambda=\End_R(M),$01 (Špenko et al., 2017). More papers on cones over reflexive polytopes, conic modules, and almost simplicial cones indicate a convergence of GIT, toric-stack, and combinatorial methods toward broader existence theorems (Malter et al., 15 Sep 2025).

A plausible implication is that the modern theory of NCCRs is no longer organized around a single construction. Instead, it is increasingly structured by equivalences among several languages—tilting bundles, modules of covariants, dimers, conic modules, exceptional collections, and mutations—each of which captures a different aspect of non-commutative crepancy.